1
00:00:00,000 --> 00:00:03,510
All the algorithms that are used in deep

2
00:00:03,510 --> 00:00:06,150
learning are largely inspired by the way

3
00:00:06,150 --> 00:00:08,880
neurons and neural networks function and

4
00:00:08,880 --> 00:00:13,200
process data in the brain. This image is

5
00:00:13,200 --> 00:00:14,969
one of the very first pictures of a

6
00:00:14,969 --> 00:00:18,539
neuron. It was drawn by Santiago Ramon

7
00:00:18,539 --> 00:00:19,170
y Cajal,

8
00:00:19,170 --> 00:00:24,029
back in 1899 based on what he saw after

9
00:00:24,029 --> 00:00:25,890
placing a pigeon's brain under the

10
00:00:25,890 --> 00:00:28,560
microscope. He is now known as the father

11
00:00:28,560 --> 00:00:31,560
of modern neuroscience, but based on his

12
00:00:31,560 --> 00:00:34,200
drawing, the neurons, one of them labeled

13
00:00:34,200 --> 00:00:37,590
A, have big bodies in the middle and long

14
00:00:37,590 --> 00:00:40,200
arms that stretch out and branch off to

15
00:00:40,200 --> 00:00:43,050
connect with other neurons. This other

16
00:00:43,050 --> 00:00:44,940
image here is that of a neural network

17
00:00:44,940 --> 00:00:47,640
and has a bunch or thousands of neurons

18
00:00:47,640 --> 00:00:50,430
in what looks like a brain tissue. It

19
00:00:50,430 --> 00:00:52,469
gives you a sense of how tightly they

20
00:00:52,469 --> 00:00:54,930
are packed together and how many of them

21
00:00:54,930 --> 00:00:59,940
are in a small brain tissue. Going back

22
00:00:59,940 --> 00:01:02,160
to the drawing of neurons by Ramon y Cajal,

23
00:01:02,160 --> 00:01:05,159
let's rotate it 90 degrees to the

24
00:01:05,159 --> 00:01:05,610
left.

25
00:01:05,610 --> 00:01:09,180
I bet this way it is starting to look a

26
00:01:09,180 --> 00:01:11,220
little familiar since it slightly

27
00:01:11,220 --> 00:01:13,350
resembles drawings of artificial neural

28
00:01:13,350 --> 00:01:18,659
networks that you must have seen. Here is

29
00:01:18,659 --> 00:01:22,020
a cartoon drawing of the neuron. The main

30
00:01:22,020 --> 00:01:24,390
body of the neuron is called the soma,

31
00:01:24,390 --> 00:01:26,990
which contains the nucleus of the neuron.

32
00:01:26,990 --> 00:01:30,240
The big network of arms sticking out of

33
00:01:30,240 --> 00:01:32,369
the body is called the dendrites, and

34
00:01:32,369 --> 00:01:35,520
then the long arm that sticks out of the

35
00:01:35,520 --> 00:01:37,740
soma in the other direction is called

36
00:01:37,740 --> 00:01:40,470
the axon. The whiskers at the end of the

37
00:01:40,470 --> 00:01:42,810
axon are called the terminal buttons or

38
00:01:42,810 --> 00:01:46,020
synapses. So the dendrites receive

39
00:01:46,020 --> 00:01:48,090
electrical impulses which carry

40
00:01:48,090 --> 00:01:51,090
information, or data, from sensors or

41
00:01:51,090 --> 00:01:52,799
terminal buttons of other adjoining

42
00:01:52,799 --> 00:01:55,560
neurons. The dendrites then carry the

43
00:01:55,560 --> 00:01:58,320
impulses or data to the soma. In the

44
00:01:58,320 --> 00:02:01,110
nucleus, electrical impulses, or the data,

45
00:02:01,110 --> 00:02:03,960
are processed by combining them together,

46
00:02:03,960 --> 00:02:06,869
and then they are passed on to the axon.

47
00:02:06,869 --> 00:02:09,690
The axon then carries the processed

48
00:02:09,690 --> 00:02:11,670
information to the terminal button or

49
00:02:11,670 --> 00:02:13,650
synapse, and the output

50
00:02:13,650 --> 00:02:15,989
of this neuron becomes the input to

51
00:02:15,989 --> 00:02:19,439
thousands of other neurons. Learning in

52
00:02:19,439 --> 00:02:21,810
the brain occurs by repeatedly

53
00:02:21,810 --> 00:02:23,970
activating certain neural connections

54
00:02:23,970 --> 00:02:26,970
over others, and this reinforces those

55
00:02:26,970 --> 00:02:29,730
connections. This makes them more likely

56
00:02:29,730 --> 00:02:32,609
to produce a desired outcome given a

57
00:02:32,609 --> 00:02:35,790
specified input. Once the desired outcome

58
00:02:35,790 --> 00:02:38,250
occurs, the neural connections causing

59
00:02:38,250 --> 00:02:43,519
that outcome become strengthened. An

60
00:02:43,849 --> 00:02:46,620
artificial neuron behaves in the same

61
00:02:46,620 --> 00:02:49,319
way as a biological neuron. So it

62
00:02:49,319 --> 00:02:53,849
consists of a soma, dendrites, and an axon

63
00:02:53,849 --> 00:02:56,940
to pass on the output of this neuron to

64
00:02:56,940 --> 00:03:00,060
other neurons. The end of the axon can

65
00:03:00,060 --> 00:03:01,680
branch off to connect to many other

66
00:03:01,680 --> 00:03:04,230
neurons, but for simplicity we are just

67
00:03:04,230 --> 00:03:06,959
showing one branch here. The learning

68
00:03:06,959 --> 00:03:09,540
process also very much resembles the way

69
00:03:09,540 --> 00:03:12,359
learning occurs in the brain as you will

70
00:03:12,359 --> 00:03:16,709
see in the next couple of videos. Now

71
00:03:16,709 --> 00:03:18,720
that we understand the different parts

72
00:03:18,720 --> 00:03:21,359
of an artificial neuron, let's learn how

73
00:03:21,359 --> 00:03:23,190
we formulate the way artificial neural

74
00:03:23,190 --> 00:03:26,190
networks process information.